Available configs

The configs/train/ directory ships 17 training presets. Most are named <model>_<gpu-count>_<parallelism>.toml so the filename doubles as a recipe. The configs/model/ directory ships 4 model-only presets used by scripts/convert_checkpoint.py for DCP ↔ HuggingFace conversion.

Dense training configs

File	Purpose	Model	GPUs	Parallelism	Notes
debug.toml	Tiny model, fast smoke test	256-dim, 4-layer	1+	FSDP (dp_shard=-1)	100 steps, compile_model=false, act_ckpt="none"
hf_wikitext.toml	End-to-end HuggingFace streaming	512-dim, 8-layer	1+	FSDP (dp_shard=-1)	GPT-2 tokenizer, wikitext-103-raw-v1, 500 steps
7b.toml	General-purpose Llama-3 7B	7B	any	FSDP (dp_shard=-1)	100K steps, compile_model=true, full AC
7b_32gpu_fsdp.toml	Baseline multi-node 7B	7B	32	pure FSDP	2M tokens/step, simplest multi-node recipe
7b_12gpu_tp4.toml	7B with intra-node TP	7B	12	TP=4 × FSDP=3	TP within node (NVLink), FSDP across (IB)
7b_16gpu_adamw.toml	Long preemptible 7B run	7B	16	FSDP (dp_shard=-1)	100K steps, 210B tokens, ckpt every 500 steps
7b_16gpu_fp8.toml	FP8 compute + FSDP2 float8 AG	7B	16	FSDP (dp_shard=-1)	mixed_precision="fp8", E4M3 fwd / E5M2 bwd
7b_16gpu_muon.toml	Muon optimizer + z-loss + chunked CE	7B	16	FSDP (dp_shard=-1)	Tests Muon, chunked cross-entropy, z-loss together
13b_24gpu_validation.toml	Full-stack validation run	13B	24	TP=4 × FSDP=6	WandB, profiling, eval, HF tokenizer, 1000 steps
13b_32gpu_tp4_pp2.toml	13B with pipeline parallel	13B	32	TP=4 × PP=2 × FSDP=4	40 layers → 20 per PP stage, 1f1b schedule
29b_32gpu_tp4_pp2.toml	Custom 29B sized for H200 140GB	29B	32	TP=4 × PP=2 × FSDP=4	dim=6144, 56 layers, saturates 120 GB/GPU
70b_32gpu_tp4.toml	70B without PP bubble	70B	32	TP=4 × FSDP=8	No PP — fits via FSDP sharding alone
70b_32gpu_tp4_pp4.toml	70B when memory is tight	70B	32	TP=4 × PP=4 × FSDP=2	80 layers → 20 per PP stage, less FSDP sharding

MoE training configs

File	Purpose	Model	GPUs	Parallelism	MoE
debug_moe.toml	Tiny MoE smoke test	256-dim, 4-layer	1+	FSDP (dp_shard=-1)	4 experts, top-2, moe_frequency=2
moe_8gpu_stress.toml	Saturate 2 nodes with MoE	~4B total / 1.8B active	8	TP=4 × FSDP=2	8 experts, top-2, moe_frequency=1
moe_24gpu.toml	24-GPU MoE stress test	~7B total / 1.8B active	24	TP=4 × FSDP=6	8 experts, top-2, grad_accum=32 for full saturation
moe_ep_32gpu.toml	MoE + Expert Parallel	~4B total / 1.8B active	32	TP=4 × EP=2 × FSDP=4	8 experts, top-2, all-to-all across IB

See the MoE Expert Parallel benchmark for the numbers the last config produced.

Model-only configs

These don’t include training fields — they’re loaded by scripts/convert_checkpoint.py to describe the architecture when round-tripping checkpoints to HuggingFace.

File	Architecture
llama_7b.toml	dim=4096, 32 layers, 32 heads, 8 kv-heads, ffn_dim_multiplier=1.3
llama_13b.toml	dim=5120, 40 layers, 40 heads, 8 kv-heads, ffn_dim_multiplier=1.3
llama_70b.toml	dim=8192, 80 layers, 64 heads, 8 kv-heads, ffn_hidden_dim=28672
moe_small.toml	dim=2048, 24 layers, 16 heads, 4 kv-heads, 8 experts top-2

Conventions

• dp_shard = -1 in a config means “fill the remaining mesh dimension with FSDP” — the loader resolves this to world_size / (dp_replicate·tp·pp·cp·ep). Most single-dimension configs (7B, FP8, Muon, debug) use -1 so the same config works on any GPU count.

• compile_model = false is set on every MoE config. Routing produces data-dependent shapes that break torch.compile’s graph — JobConfig.validate(world_size) logs a warning (not an error) if you combine them.

• Paths in these configs (dataset_path = "/path/to/...") are placeholders — replace them with a real tokenized shard directory before running. The hf_wikitext.toml config is the only one that runs end-to-end without path edits (it streams from the HF Hub).

• Short max_steps (20–100) in the multi-node configs is a benchmark-sizing default, not a training budget. Override with --train.max_steps=… for real runs.

See also

• Parallelism recipes — same data, indexed by (model, GPU count) rather than by filename.

• Benchmarks — measured throughput for the configs that were benchmarked end-to-end.

• Config sections — the fields every TOML key maps to.